Poynter: how fact-checkers are dealing with AI's reliability problems

Written by Louis Jacobson and published June 22, 2026, this Poynter piece covers conversations from GlobalFact—the annual conference for fact-checking organizations—about the practical challenges AI tools create for journalists whose job is to verify what is true.

The context is that fact-checkers are among the professionals most exposed to AI’s accuracy failures, because inaccurate AI output is exactly what they are trained to detect. The article collects observations from people who have tested AI chatbots systematically and run controlled comparisons across models.

One finding that comes up in the reporting is that AI models can seem highly reliable on simple factual questions while generating confident misinformation on less common topics. The gap between apparent reliability and actual reliability is a practical problem for newsrooms that want to use AI in research workflows without inadvertently laundering hallucinations into published work.

A second area the article covers is ideological bias. Researchers who compared ChatGPT, DeepSeek, and other models found that responses tend to reflect the ideological structures and data sources embedded in each model’s training. Models from different national or commercial contexts produce systematically different outputs on politically contested topics—not randomly wrong, but consistently skewed in directions that correlate with their origins.

A third concern is deepfakes. Experts quoted in the piece note that AI-generated false content does not need to be undetectable to spread. Material that is plausible enough to circulate faster than corrections can follow is already effective disinformation, regardless of whether sophisticated analysis could identify it as generated.

The article does not provide specific workflow guidance for integrating AI into fact-checking. Its value is diagnostic rather than prescriptive: it documents where the real reliability problems are, which is a prerequisite for designing trustworthy AI-assisted workflows in editorial environments.